fcn

FCN

Fully Convolutional Networks for Semantic Segmentation CVPR 2015

Method Overview

Fully Convolutional Network (FCN) was the most classic semantic segmentation network, its formulation is even before the ResNet era. Our implementation of FCN borrows from TorchVision, which is a modern take of FCN with ResNet backbone. The "replacing stride by dilation" option from Deeplab is integrated to maintain a 8x down-sampled feature map that makes prediction through a simple segmentation head. It can still be seen as the most simple segmentation baseline, but it is not the original FCN anymore.

Results

Training time estimated with single 2080 Ti.

ImageNet pre-training, 3-times average/best.

PASCAL VOC 2012 trainaug (val)

backbone	resolution	training time	precision	mIoU (avg)	mIoU
ResNet-101	321 x 321	3.3h	mix	70.72	70.83	model | shell
ResNet-101	321 x 321	6.3h	full	70.91	71.55	model | shell

Cityscapes (val)

backbone	resolution	training time	precision	mIoU (avg)	mIoU
ResNet-101	321 x 321	2.2h	mix	68.05	68.20	model | shell

Profiling

FPS is best trial-avg among 3 trials on a 2080 Ti.

backbone	resolution	FPS	FLOPS(G)	Params(M)
ResNet-101	256 x 512	43.32	216.42	51.95
ResNet-101	512 x 1024	12.06	865.69	51.95
ResNet-101	1024 x 2048	3.06	3462.77	51.95

Citation

@inproceedings{long2015fully,
  title={Fully convolutional networks for semantic segmentation},
  author={Long, Jonathan and Shelhamer, Evan and Darrell, Trevor},
  booktitle={Computer Vision and Pattern Recognition},
  year={2015}
}

@article{shelhamer2016fully,
  title={Fully convolutional networks for semantic segmentation},
  author={Shelhamer, Evan and Long, Jonathan and Darrell, Trevor},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={39},
  number={4},
  pages={640--651},
  year={2016}
}